Schrödinger’s AI Bubble

A Saturday Edition of ABP

While we spend most of our time in Hard Tech, we can’t ignore what’s going on in AI given how much of the current capital markets are driven by it. Here’s our current attempt to contextualize today’s AI opportunity:

AI in 2025 feels like Schrödinger’s bubble. It’s both overinflated and undervalued at the same time. But when one stops treating “AI” as a single market, the contradiction starts to make a lot more sense. Let’s break it down:

The frontier-model world is easy to call a bubble. It has the familiar shine of too much capital chasing too few differentiated ideas. Training budget growth is outpacing revenue in heroic fashion. GPU clusters are being bought faster than customers are being signed. And many teams are shipping similar architectures and competing mostly on scale, marketing, or subsidies.

But if one looks at the real economy, the story flips. Sectors like manufacturing, logistics, agriculture, energy, construction, healthcare, and defense (what we refer to collectively as the “Physical Enterprise”) are still in the earliest stages of AI adoption. The gap between what’s technically possible and what’s deployed is enormous. The biggest opportunities are hiding in workflows that haven’t been touched yet. The catch here is that the GTM for these opportunities may be slower to show revenue growth, and will often look like services that then productize over time.

Given this, trying to label all of AI a bubble or not a bubble misses the structure of what’s actually happening. AI is both. Speculative at the top, underdeveloped at the edges.

In the long arc, AI undoubtedly will be the largest productivity driver of our time. For founders and investors, the opportunity is figuring out which parts of it are mispriced and when to lean into those places where the market hasn’t looked yet.